Z.ai, the Chinese AI laboratory, has released GLM-5.1, a 754-billion-parameter open-source large language model designed to operate autonomously on a single task for up to eight hours, pushing the boundaries of what AI agents can accomplish without human input.
The model's weights are publicly available on Hugging Face and ModelScope, the AI model hosting platforms, and it can be deployed locally via vLLM, SGLang and xLLM.
The release follows a proprietary GLM-5 Turbo variant and reflects Z.ai's strategy of pairing open-source flagship models with commercial, optimised releases designed to generate subscription and API revenue.
Z.ai describes the model's core technical advance as a staircase pattern of optimisation, in which the system makes sustained, incremental progress across long task sequences rather than plateauing early.
In one benchmark scenario using VectorDBBench, a database performance testing suite, the model ran 655 iterations and more than 6,000 tool calls over the course of a single task, improving throughput from an initial 3,547 queries per second to 21,500 queries per second after six rounds of structural improvement.
On KernelBench Level 3, a test of low-level code optimisation, GLM-5.1 achieved a 3.6x geometric mean speedup across 50 problems, ahead of the earlier GLM-5's plateau of 2.6x, though behind Claude Opus 4.6's 4.2x result on the same test.
The model posted a score of 58.4 on SWE-Bench Pro, a benchmark measuring AI performance on real-world software engineering tasks, narrowly ahead of GPT-5.4 at 57.7 and Claude Opus 4.6 at 57.3.
Z.ai said GLM-5.1 also recorded gains on Terminal-Bench 2.0, CyberGym, MCP-Atlas and T3-Bench, though independent verification of those figures was not immediately available.
Z.ai's leader, known on X as Lou, said autonomous work duration may prove as significant a development curve for AI as scaling laws, the relationship between model size and capability that has driven progress in the field for years.
Related reading
- Memento-Skills lets AI agents rewrite their own capabilities without retraining underlying models
- Anthropic cuts off Claude Code subscribers from third-party AI agents, starting with OpenClaw
- Microsoft launches three AI models under new superintelligence research unit
Lou noted that AI agents capable of around 20 steps at the end of last year can now, with GLM-5.1, execute up to 1,700 steps, adding that the open-source community would be able to verify the claim directly.
Z.ai has published subscription tiers and API pricing alongside the release, including quarter-based coding plans and per-token rates, with a promotional period offering reduced off-peak billing.
The recap
- Z.ai releases GLM-5.1 under MIT License, weights on Hugging Face
- Model has 754 billion parameters and 202,752 token context window
- Commercial tiers and API pricing announced, local deployment supported
